fixup! feat(core.utils): use cuda.bindings.nvvm.version() in NVVM fingerprint

Author: cpcloud

cuda_core/cuda/core/utils/_program_cache.py

@@ -344,17 +344,20 @@ def _linker_backend_and_version() -> tuple[str, str]:
 def _nvvm_fingerprint() -> str:
     """Stable identifier for the loaded NVVM toolchain.
 
-    NVVM lacks a direct version API (nvbugs 5312315), but ``ir_version()``
-    reports the IR major/minor/debug pair the toolchain emits -- enough to
-    keep pre-/post-upgrade caches separate. Paired with the driver and
-    cuda-core versions already in the digest, this is a practical substitute
-    for a true libNVVM version.
+    Combines the libNVVM library version (``cuda.bindings.nvvm.version()``)
+    with the IR version reported by ``ir_version()``. The library version is
+    the primary invalidation lever: a libNVVM patch upgrade can change
+    codegen while keeping the same IR major/minor, so keying only on the IR
+    pair would silently reuse stale entries. Paired with cuda-core, the IR
+    pair adds defence in depth without making the key any less stable.
     """
+    from cuda.bindings import nvvm
     from cuda.core._program import _get_nvvm_module
 
-    nvvm = _get_nvvm_module()
-    major, minor, debug_major, debug_minor = nvvm.ir_version()
-    return f"ir={major}.{minor}.{debug_major}.{debug_minor}"
+    module = _get_nvvm_module()
+    lib_major, lib_minor = nvvm.version()
+    major, minor, debug_major, debug_minor = module.ir_version()
+    return f"lib={lib_major}.{lib_minor};ir={major}.{minor}.{debug_major}.{debug_minor}"
 
 
 def _cuda_core_version() -> str:
@@ -1083,8 +1086,11 @@ class FileStreamProgramCache(ProgramCacheResource):
         ``cuda.core`` patch releases because every entry's key encodes
         the relevant backend/compiler/runtime fingerprints for its
         compilation path (NVRTC entries pin the NVRTC version, NVVM
-        entries pin the NVVM IR version, PTX/linker entries pin the
-        chosen linker backend, its version, and the driver version).
+        entries pin the libNVVM library and IR versions, PTX/linker
+        entries pin the chosen linker backend and its version -- and,
+        when the cuLink/driver backend is selected, the driver version
+        too; nvJitLink-backed PTX entries are deliberately driver-version
+        independent).
 
     Parameters
     ----------

cuda_core/tests/test_program_cache.py

@@ -503,6 +503,20 @@ def test_make_program_cache_key_nvvm_use_libdevice_false_equals_none():
     assert k_true != k_none
 
 
+def test_make_program_cache_key_nvvm_library_version_changes_key(monkeypatch):
+    """Updating libNVVM (different ``cuda.bindings.nvvm.version()``) must
+    invalidate NVVM cache entries -- a patch upgrade can change codegen
+    without bumping the IR version, so keying only on IR would silently
+    return stale entries."""
+    from cuda.core.utils import _program_cache
+
+    monkeypatch.setattr(_program_cache, "_nvvm_fingerprint", lambda: "lib=12.3;ir=1.8.3.0")
+    k_old = _make_key(code="abc", code_type="nvvm", target_type="ptx")
+    monkeypatch.setattr(_program_cache, "_nvvm_fingerprint", lambda: "lib=12.4;ir=1.8.3.0")
+    k_new = _make_key(code="abc", code_type="nvvm", target_type="ptx")
+    assert k_old != k_new
+
+
 def test_make_program_cache_key_nvvm_probe_changes_key(monkeypatch):
     """NVVM keys must reflect the NVVM toolchain identity (IR version)
     so an upgraded libNVVM does not silently reuse pre-upgrade entries."""